hw/vfio/common.c

   1 /*
   2  * generic functions used by VFIO devices
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Alex Williamson <alex.williamson@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Based on qemu-kvm device-assignment:
  13  *  Adapted for KVM by Qumranet.
  14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
  15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
  16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
  17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
  18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
  19  */
  20
  21 #include <sys/ioctl.h>
  22 #include <sys/mman.h>
  23 #include <linux/vfio.h>
  24
  25 #include "hw/vfio/vfio-common.h"
  26 #include "hw/vfio/vfio.h"
  27 #include "exec/address-spaces.h"
  28 #include "exec/memory.h"
  29 #include "hw/hw.h"
  30 #include "qemu/error-report.h"
  31 #include "sysemu/kvm.h"
  32 #include "trace.h"
  33
  34 struct vfio_group_head vfio_group_list =
  35     QLIST_HEAD_INITIALIZER(vfio_group_list);
  36 struct vfio_as_head vfio_address_spaces =
  37     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
  38
  39 #ifdef CONFIG_KVM
  40 /*
  41  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
  42  * for the life of the VM.  Closing the file descriptor only drops our
  43  * reference to it and the device's reference to kvm.  Therefore once
  44  * initialized, this file descriptor is only released on QEMU exit and
  45  * we'll re-use it should another vfio device be attached before then.
  46  */
  47 static int vfio_kvm_device_fd = -1;
  48 #endif
  49
  50 /*
  51  * Common VFIO interrupt disable
  52  */
  53 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
  54 {
  55     struct vfio_irq_set irq_set = {
  56         .argsz = sizeof(irq_set),
  57         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
  58         .index = index,
  59         .start = 0,
  60         .count = 0,
  61     };
  62
  63     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  64 }
  65
  66 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
  67 {
  68     struct vfio_irq_set irq_set = {
  69         .argsz = sizeof(irq_set),
  70         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
  71         .index = index,
  72         .start = 0,
  73         .count = 1,
  74     };
  75
  76     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  77 }
  78
  79 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
  80 {
  81     struct vfio_irq_set irq_set = {
  82         .argsz = sizeof(irq_set),
  83         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
  84         .index = index,
  85         .start = 0,
  86         .count = 1,
  87     };
  88
  89     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  90 }
  91
  92 /*
  93  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
  94  */
  95 void vfio_region_write(void *opaque, hwaddr addr,
  96                        uint64_t data, unsigned size)
  97 {
  98     VFIORegion *region = opaque;
  99     VFIODevice *vbasedev = region->vbasedev;
 100     union {
 101         uint8_t byte;
 102         uint16_t word;
 103         uint32_t dword;
 104         uint64_t qword;
 105     } buf;
 106
 107     switch (size) {
 108     case 1:
 109         buf.byte = data;
 110         break;
 111     case 2:
 112         buf.word = cpu_to_le16(data);
 113         break;
 114     case 4:
 115         buf.dword = cpu_to_le32(data);
 116         break;
 117     default:
 118         hw_error("vfio: unsupported write size, %d bytes", size);
 119         break;
 120     }
 121
 122     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 123         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
 124                      ",%d) failed: %m",
 125                      __func__, vbasedev->name, region->nr,
 126                      addr, data, size);
 127     }
 128
 129     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
 130
 131     /*
 132      * A read or write to a BAR always signals an INTx EOI.  This will
 133      * do nothing if not pending (including not in INTx mode).  We assume
 134      * that a BAR access is in response to an interrupt and that BAR
 135      * accesses will service the interrupt.  Unfortunately, we don't know
 136      * which access will service the interrupt, so we're potentially
 137      * getting quite a few host interrupts per guest interrupt.
 138      */
 139     vbasedev->ops->vfio_eoi(vbasedev);
 140 }
 141
 142 uint64_t vfio_region_read(void *opaque,
 143                           hwaddr addr, unsigned size)
 144 {
 145     VFIORegion *region = opaque;
 146     VFIODevice *vbasedev = region->vbasedev;
 147     union {
 148         uint8_t byte;
 149         uint16_t word;
 150         uint32_t dword;
 151         uint64_t qword;
 152     } buf;
 153     uint64_t data = 0;
 154
 155     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 156         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
 157                      __func__, vbasedev->name, region->nr,
 158                      addr, size);
 159         return (uint64_t)-1;
 160     }
 161     switch (size) {
 162     case 1:
 163         data = buf.byte;
 164         break;
 165     case 2:
 166         data = le16_to_cpu(buf.word);
 167         break;
 168     case 4:
 169         data = le32_to_cpu(buf.dword);
 170         break;
 171     default:
 172         hw_error("vfio: unsupported read size, %d bytes", size);
 173         break;
 174     }
 175
 176     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
 177
 178     /* Same as write above */
 179     vbasedev->ops->vfio_eoi(vbasedev);
 180
 181     return data;
 182 }
 183
 184 const MemoryRegionOps vfio_region_ops = {
 185     .read = vfio_region_read,
 186     .write = vfio_region_write,
 187     .endianness = DEVICE_LITTLE_ENDIAN,
 188 };
 189
 190 /*
 191  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
 192  */
 193 static int vfio_dma_unmap(VFIOContainer *container,
 194                           hwaddr iova, ram_addr_t size)
 195 {
 196     struct vfio_iommu_type1_dma_unmap unmap = {
 197         .argsz = sizeof(unmap),
 198         .flags = 0,
 199         .iova = iova,
 200         .size = size,
 201     };
 202
 203     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
 204         error_report("VFIO_UNMAP_DMA: %d\n", -errno);
 205         return -errno;
 206     }
 207
 208     return 0;
 209 }
 210
 211 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
 212                         ram_addr_t size, void *vaddr, bool readonly)
 213 {
 214     struct vfio_iommu_type1_dma_map map = {
 215         .argsz = sizeof(map),
 216         .flags = VFIO_DMA_MAP_FLAG_READ,
 217         .vaddr = (__u64)(uintptr_t)vaddr,
 218         .iova = iova,
 219         .size = size,
 220     };
 221
 222     if (!readonly) {
 223         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
 224     }
 225
 226     /*
 227      * Try the mapping, if it fails with EBUSY, unmap the region and try
 228      * again.  This shouldn't be necessary, but we sometimes see it in
 229      * the the VGA ROM space.
 230      */
 231     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
 232         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
 233          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
 234         return 0;
 235     }
 236
 237     error_report("VFIO_MAP_DMA: %d\n", -errno);
 238     return -errno;
 239 }
 240
 241 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 242 {
 243     return (!memory_region_is_ram(section->mr) &&
 244             !memory_region_is_iommu(section->mr)) ||
 245            /*
 246             * Sizing an enabled 64-bit BAR can cause spurious mappings to
 247             * addresses in the upper part of the 64-bit address space.  These
 248             * are never accessed by the CPU and beyond the address width of
 249             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
 250             */
 251            section->offset_within_address_space & (1ULL << 63);
 252 }
 253
 254 static void vfio_iommu_map_notify(Notifier *n, void *data)
 255 {
 256     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
 257     VFIOContainer *container = giommu->container;
 258     IOMMUTLBEntry *iotlb = data;
 259     MemoryRegion *mr;
 260     hwaddr xlat;
 261     hwaddr len = iotlb->addr_mask + 1;
 262     void *vaddr;
 263     int ret;
 264
 265     trace_vfio_iommu_map_notify(iotlb->iova,
 266                                 iotlb->iova + iotlb->addr_mask);
 267
 268     /*
 269      * The IOMMU TLB entry we have just covers translation through
 270      * this IOMMU to its immediate target.  We need to translate
 271      * it the rest of the way through to memory.
 272      */
 273     mr = address_space_translate(&address_space_memory,
 274                                  iotlb->translated_addr,
 275                                  &xlat, &len, iotlb->perm & IOMMU_WO);
 276     if (!memory_region_is_ram(mr)) {
 277         error_report("iommu map to non memory area %"HWADDR_PRIx"\n",
 278                      xlat);
 279         return;
 280     }
 281     /*
 282      * Translation truncates length to the IOMMU page size,
 283      * check that it did not truncate too much.
 284      */
 285     if (len & iotlb->addr_mask) {
 286         error_report("iommu has granularity incompatible with target AS\n");
 287         return;
 288     }
 289
 290     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
 291         vaddr = memory_region_get_ram_ptr(mr) + xlat;
 292         ret = vfio_dma_map(container, iotlb->iova,
 293                            iotlb->addr_mask + 1, vaddr,
 294                            !(iotlb->perm & IOMMU_WO) || mr->readonly);
 295         if (ret) {
 296             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 297                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
 298                          container, iotlb->iova,
 299                          iotlb->addr_mask + 1, vaddr, ret);
 300         }
 301     } else {
 302         ret = vfio_dma_unmap(container, iotlb->iova, iotlb->addr_mask + 1);
 303         if (ret) {
 304             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 305                          "0x%"HWADDR_PRIx") = %d (%m)",
 306                          container, iotlb->iova,
 307                          iotlb->addr_mask + 1, ret);
 308         }
 309     }
 310 }
 311
 312 static void vfio_listener_region_add(MemoryListener *listener,
 313                                      MemoryRegionSection *section)
 314 {
 315     VFIOContainer *container = container_of(listener, VFIOContainer,
 316                                             iommu_data.type1.listener);
 317     hwaddr iova, end;
 318     Int128 llend;
 319     void *vaddr;
 320     int ret;
 321
 322     if (vfio_listener_skipped_section(section)) {
 323         trace_vfio_listener_region_add_skip(
 324                 section->offset_within_address_space,
 325                 section->offset_within_address_space +
 326                 int128_get64(int128_sub(section->size, int128_one())));
 327         return;
 328     }
 329
 330     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 331                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 332         error_report("%s received unaligned region", __func__);
 333         return;
 334     }
 335
 336     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 337     llend = int128_make64(section->offset_within_address_space);
 338     llend = int128_add(llend, section->size);
 339     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 340
 341     if (int128_ge(int128_make64(iova), llend)) {
 342         return;
 343     }
 344
 345     memory_region_ref(section->mr);
 346
 347     if (memory_region_is_iommu(section->mr)) {
 348         VFIOGuestIOMMU *giommu;
 349
 350         trace_vfio_listener_region_add_iommu(iova,
 351                     int128_get64(int128_sub(llend, int128_one())));
 352         /*
 353          * FIXME: We should do some checking to see if the
 354          * capabilities of the host VFIO IOMMU are adequate to model
 355          * the guest IOMMU
 356          *
 357          * FIXME: For VFIO iommu types which have KVM acceleration to
 358          * avoid bouncing all map/unmaps through qemu this way, this
 359          * would be the right place to wire that up (tell the KVM
 360          * device emulation the VFIO iommu handles to use).
 361          */
 362         /*
 363          * This assumes that the guest IOMMU is empty of
 364          * mappings at this point.
 365          *
 366          * One way of doing this is:
 367          * 1. Avoid sharing IOMMUs between emulated devices or different
 368          * IOMMU groups.
 369          * 2. Implement VFIO_IOMMU_ENABLE in the host kernel to fail if
 370          * there are some mappings in IOMMU.
 371          *
 372          * VFIO on SPAPR does that. Other IOMMU models may do that different,
 373          * they must make sure there are no existing mappings or
 374          * loop through existing mappings to map them into VFIO.
 375          */
 376         giommu = g_malloc0(sizeof(*giommu));
 377         giommu->iommu = section->mr;
 378         giommu->container = container;
 379         giommu->n.notify = vfio_iommu_map_notify;
 380         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
 381         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
 382
 383         return;
 384     }
 385
 386     /* Here we assume that memory_region_is_ram(section->mr)==true */
 387
 388     end = int128_get64(llend);
 389     vaddr = memory_region_get_ram_ptr(section->mr) +
 390             section->offset_within_region +
 391             (iova - section->offset_within_address_space);
 392
 393     trace_vfio_listener_region_add_ram(iova, end - 1, vaddr);
 394
 395     ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
 396     if (ret) {
 397         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 398                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
 399                      container, iova, end - iova, vaddr, ret);
 400
 401         /*
 402          * On the initfn path, store the first error in the container so we
 403          * can gracefully fail.  Runtime, there's not much we can do other
 404          * than throw a hardware error.
 405          */
 406         if (!container->iommu_data.type1.initialized) {
 407             if (!container->iommu_data.type1.error) {
 408                 container->iommu_data.type1.error = ret;
 409             }
 410         } else {
 411             hw_error("vfio: DMA mapping failed, unable to continue");
 412         }
 413     }
 414 }
 415
 416 static void vfio_listener_region_del(MemoryListener *listener,
 417                                      MemoryRegionSection *section)
 418 {
 419     VFIOContainer *container = container_of(listener, VFIOContainer,
 420                                             iommu_data.type1.listener);
 421     hwaddr iova, end;
 422     int ret;
 423
 424     if (vfio_listener_skipped_section(section)) {
 425         trace_vfio_listener_region_del_skip(
 426                 section->offset_within_address_space,
 427                 section->offset_within_address_space +
 428                 int128_get64(int128_sub(section->size, int128_one())));
 429         return;
 430     }
 431
 432     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 433                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 434         error_report("%s received unaligned region", __func__);
 435         return;
 436     }
 437
 438     if (memory_region_is_iommu(section->mr)) {
 439         VFIOGuestIOMMU *giommu;
 440
 441         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
 442             if (giommu->iommu == section->mr) {
 443                 memory_region_unregister_iommu_notifier(&giommu->n);
 444                 QLIST_REMOVE(giommu, giommu_next);
 445                 g_free(giommu);
 446                 break;
 447             }
 448         }
 449
 450         /*
 451          * FIXME: We assume the one big unmap below is adequate to
 452          * remove any individual page mappings in the IOMMU which
 453          * might have been copied into VFIO. This works for a page table
 454          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
 455          * That may not be true for all IOMMU types.
 456          */
 457     }
 458
 459     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 460     end = (section->offset_within_address_space + int128_get64(section->size)) &
 461           TARGET_PAGE_MASK;
 462
 463     if (iova >= end) {
 464         return;
 465     }
 466
 467     trace_vfio_listener_region_del(iova, end - 1);
 468
 469     ret = vfio_dma_unmap(container, iova, end - iova);
 470     memory_region_unref(section->mr);
 471     if (ret) {
 472         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 473                      "0x%"HWADDR_PRIx") = %d (%m)",
 474                      container, iova, end - iova, ret);
 475     }
 476 }
 477
 478 static const MemoryListener vfio_memory_listener = {
 479     .region_add = vfio_listener_region_add,
 480     .region_del = vfio_listener_region_del,
 481 };
 482
 483 static void vfio_listener_release(VFIOContainer *container)
 484 {
 485     memory_listener_unregister(&container->iommu_data.type1.listener);
 486 }
 487
 488 int vfio_mmap_region(Object *obj, VFIORegion *region,
 489                      MemoryRegion *mem, MemoryRegion *submem,
 490                      void **map, size_t size, off_t offset,
 491                      const char *name)
 492 {
 493     int ret = 0;
 494     VFIODevice *vbasedev = region->vbasedev;
 495
 496     if (VFIO_ALLOW_MMAP && size && region->flags &
 497         VFIO_REGION_INFO_FLAG_MMAP) {
 498         int prot = 0;
 499
 500         if (region->flags & VFIO_REGION_INFO_FLAG_READ) {
 501             prot |= PROT_READ;
 502         }
 503
 504         if (region->flags & VFIO_REGION_INFO_FLAG_WRITE) {
 505             prot |= PROT_WRITE;
 506         }
 507
 508         *map = mmap(NULL, size, prot, MAP_SHARED,
 509                     vbasedev->fd,
 510                     region->fd_offset + offset);
 511         if (*map == MAP_FAILED) {
 512             *map = NULL;
 513             ret = -errno;
 514             goto empty_region;
 515         }
 516
 517         memory_region_init_ram_ptr(submem, obj, name, size, *map);
 518         memory_region_set_skip_dump(submem);
 519     } else {
 520 empty_region:
 521         /* Create a zero sized sub-region to make cleanup easy. */
 522         memory_region_init(submem, obj, name, 0);
 523     }
 524
 525     memory_region_add_subregion(mem, offset, submem);
 526
 527     return ret;
 528 }
 529
 530 void vfio_reset_handler(void *opaque)
 531 {
 532     VFIOGroup *group;
 533     VFIODevice *vbasedev;
 534
 535     QLIST_FOREACH(group, &vfio_group_list, next) {
 536         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 537             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
 538         }
 539     }
 540
 541     QLIST_FOREACH(group, &vfio_group_list, next) {
 542         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 543             if (vbasedev->needs_reset) {
 544                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
 545             }
 546         }
 547     }
 548 }
 549
 550 static void vfio_kvm_device_add_group(VFIOGroup *group)
 551 {
 552 #ifdef CONFIG_KVM
 553     struct kvm_device_attr attr = {
 554         .group = KVM_DEV_VFIO_GROUP,
 555         .attr = KVM_DEV_VFIO_GROUP_ADD,
 556         .addr = (uint64_t)(unsigned long)&group->fd,
 557     };
 558
 559     if (!kvm_enabled()) {
 560         return;
 561     }
 562
 563     if (vfio_kvm_device_fd < 0) {
 564         struct kvm_create_device cd = {
 565             .type = KVM_DEV_TYPE_VFIO,
 566         };
 567
 568         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
 569             error_report("Failed to create KVM VFIO device: %m\n");
 570             return;
 571         }
 572
 573         vfio_kvm_device_fd = cd.fd;
 574     }
 575
 576     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 577         error_report("Failed to add group %d to KVM VFIO device: %m",
 578                      group->groupid);
 579     }
 580 #endif
 581 }
 582
 583 static void vfio_kvm_device_del_group(VFIOGroup *group)
 584 {
 585 #ifdef CONFIG_KVM
 586     struct kvm_device_attr attr = {
 587         .group = KVM_DEV_VFIO_GROUP,
 588         .attr = KVM_DEV_VFIO_GROUP_DEL,
 589         .addr = (uint64_t)(unsigned long)&group->fd,
 590     };
 591
 592     if (vfio_kvm_device_fd < 0) {
 593         return;
 594     }
 595
 596     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 597         error_report("Failed to remove group %d from KVM VFIO device: %m",
 598                      group->groupid);
 599     }
 600 #endif
 601 }
 602
 603 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
 604 {
 605     VFIOAddressSpace *space;
 606
 607     QLIST_FOREACH(space, &vfio_address_spaces, list) {
 608         if (space->as == as) {
 609             return space;
 610         }
 611     }
 612
 613     /* No suitable VFIOAddressSpace, create a new one */
 614     space = g_malloc0(sizeof(*space));
 615     space->as = as;
 616     QLIST_INIT(&space->containers);
 617
 618     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
 619
 620     return space;
 621 }
 622
 623 static void vfio_put_address_space(VFIOAddressSpace *space)
 624 {
 625     if (QLIST_EMPTY(&space->containers)) {
 626         QLIST_REMOVE(space, list);
 627         g_free(space);
 628     }
 629 }
 630
 631 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
 632 {
 633     VFIOContainer *container;
 634     int ret, fd;
 635     VFIOAddressSpace *space;
 636
 637     space = vfio_get_address_space(as);
 638
 639     QLIST_FOREACH(container, &space->containers, next) {
 640         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
 641             group->container = container;
 642             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 643             return 0;
 644         }
 645     }
 646
 647     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
 648     if (fd < 0) {
 649         error_report("vfio: failed to open /dev/vfio/vfio: %m");
 650         ret = -errno;
 651         goto put_space_exit;
 652     }
 653
 654     ret = ioctl(fd, VFIO_GET_API_VERSION);
 655     if (ret != VFIO_API_VERSION) {
 656         error_report("vfio: supported vfio version: %d, "
 657                      "reported version: %d", VFIO_API_VERSION, ret);
 658         ret = -EINVAL;
 659         goto close_fd_exit;
 660     }
 661
 662     container = g_malloc0(sizeof(*container));
 663     container->space = space;
 664     container->fd = fd;
 665     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
 666         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
 667         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
 668
 669         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 670         if (ret) {
 671             error_report("vfio: failed to set group container: %m");
 672             ret = -errno;
 673             goto free_container_exit;
 674         }
 675
 676         ret = ioctl(fd, VFIO_SET_IOMMU,
 677                     v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU);
 678         if (ret) {
 679             error_report("vfio: failed to set iommu for container: %m");
 680             ret = -errno;
 681             goto free_container_exit;
 682         }
 683
 684         container->iommu_data.type1.listener = vfio_memory_listener;
 685         container->iommu_data.release = vfio_listener_release;
 686
 687         memory_listener_register(&container->iommu_data.type1.listener,
 688                                  container->space->as);
 689
 690         if (container->iommu_data.type1.error) {
 691             ret = container->iommu_data.type1.error;
 692             error_report("vfio: memory listener initialization failed for container");
 693             goto listener_release_exit;
 694         }
 695
 696         container->iommu_data.type1.initialized = true;
 697
 698     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) {
 699         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 700         if (ret) {
 701             error_report("vfio: failed to set group container: %m");
 702             ret = -errno;
 703             goto free_container_exit;
 704         }
 705         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU);
 706         if (ret) {
 707             error_report("vfio: failed to set iommu for container: %m");
 708             ret = -errno;
 709             goto free_container_exit;
 710         }
 711
 712         /*
 713          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
 714          * when container fd is closed so we do not call it explicitly
 715          * in this file.
 716          */
 717         ret = ioctl(fd, VFIO_IOMMU_ENABLE);
 718         if (ret) {
 719             error_report("vfio: failed to enable container: %m");
 720             ret = -errno;
 721             goto free_container_exit;
 722         }
 723
 724         container->iommu_data.type1.listener = vfio_memory_listener;
 725         container->iommu_data.release = vfio_listener_release;
 726
 727         memory_listener_register(&container->iommu_data.type1.listener,
 728                                  container->space->as);
 729
 730     } else {
 731         error_report("vfio: No available IOMMU models");
 732         ret = -EINVAL;
 733         goto free_container_exit;
 734     }
 735
 736     QLIST_INIT(&container->group_list);
 737     QLIST_INSERT_HEAD(&space->containers, container, next);
 738
 739     group->container = container;
 740     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 741
 742     return 0;
 743 listener_release_exit:
 744     vfio_listener_release(container);
 745
 746 free_container_exit:
 747     g_free(container);
 748
 749 close_fd_exit:
 750     close(fd);
 751
 752 put_space_exit:
 753     vfio_put_address_space(space);
 754
 755     return ret;
 756 }
 757
 758 static void vfio_disconnect_container(VFIOGroup *group)
 759 {
 760     VFIOContainer *container = group->container;
 761
 762     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
 763         error_report("vfio: error disconnecting group %d from container",
 764                      group->groupid);
 765     }
 766
 767     QLIST_REMOVE(group, container_next);
 768     group->container = NULL;
 769
 770     if (QLIST_EMPTY(&container->group_list)) {
 771         VFIOAddressSpace *space = container->space;
 772
 773         if (container->iommu_data.release) {
 774             container->iommu_data.release(container);
 775         }
 776         QLIST_REMOVE(container, next);
 777         trace_vfio_disconnect_container(container->fd);
 778         close(container->fd);
 779         g_free(container);
 780
 781         vfio_put_address_space(space);
 782     }
 783 }
 784
 785 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
 786 {
 787     VFIOGroup *group;
 788     char path[32];
 789     struct vfio_group_status status = { .argsz = sizeof(status) };
 790
 791     QLIST_FOREACH(group, &vfio_group_list, next) {
 792         if (group->groupid == groupid) {
 793             /* Found it.  Now is it already in the right context? */
 794             if (group->container->space->as == as) {
 795                 return group;
 796             } else {
 797                 error_report("vfio: group %d used in multiple address spaces",
 798                              group->groupid);
 799                 return NULL;
 800             }
 801         }
 802     }
 803
 804     group = g_malloc0(sizeof(*group));
 805
 806     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
 807     group->fd = qemu_open(path, O_RDWR);
 808     if (group->fd < 0) {
 809         error_report("vfio: error opening %s: %m", path);
 810         goto free_group_exit;
 811     }
 812
 813     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
 814         error_report("vfio: error getting group status: %m");
 815         goto close_fd_exit;
 816     }
 817
 818     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
 819         error_report("vfio: error, group %d is not viable, please ensure "
 820                      "all devices within the iommu_group are bound to their "
 821                      "vfio bus driver.", groupid);
 822         goto close_fd_exit;
 823     }
 824
 825     group->groupid = groupid;
 826     QLIST_INIT(&group->device_list);
 827
 828     if (vfio_connect_container(group, as)) {
 829         error_report("vfio: failed to setup container for group %d", groupid);
 830         goto close_fd_exit;
 831     }
 832
 833     if (QLIST_EMPTY(&vfio_group_list)) {
 834         qemu_register_reset(vfio_reset_handler, NULL);
 835     }
 836
 837     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
 838
 839     vfio_kvm_device_add_group(group);
 840
 841     return group;
 842
 843 close_fd_exit:
 844     close(group->fd);
 845
 846 free_group_exit:
 847     g_free(group);
 848
 849     return NULL;
 850 }
 851
 852 void vfio_put_group(VFIOGroup *group)
 853 {
 854     if (!group || !QLIST_EMPTY(&group->device_list)) {
 855         return;
 856     }
 857
 858     vfio_kvm_device_del_group(group);
 859     vfio_disconnect_container(group);
 860     QLIST_REMOVE(group, next);
 861     trace_vfio_put_group(group->fd);
 862     close(group->fd);
 863     g_free(group);
 864
 865     if (QLIST_EMPTY(&vfio_group_list)) {
 866         qemu_unregister_reset(vfio_reset_handler, NULL);
 867     }
 868 }
 869
 870 int vfio_get_device(VFIOGroup *group, const char *name,
 871                        VFIODevice *vbasedev)
 872 {
 873     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
 874     int ret, fd;
 875
 876     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
 877     if (fd < 0) {
 878         error_report("vfio: error getting device %s from group %d: %m",
 879                      name, group->groupid);
 880         error_printf("Verify all devices in group %d are bound to vfio-<bus> "
 881                      "or pci-stub and not already in use\n", group->groupid);
 882         return fd;
 883     }
 884
 885     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
 886     if (ret) {
 887         error_report("vfio: error getting device info: %m");
 888         close(fd);
 889         return ret;
 890     }
 891
 892     vbasedev->fd = fd;
 893     vbasedev->group = group;
 894     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
 895
 896     vbasedev->num_irqs = dev_info.num_irqs;
 897     vbasedev->num_regions = dev_info.num_regions;
 898     vbasedev->flags = dev_info.flags;
 899
 900     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
 901                           dev_info.num_irqs);
 902
 903     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
 904     return 0;
 905 }
 906
 907 void vfio_put_base_device(VFIODevice *vbasedev)
 908 {
 909     if (!vbasedev->group) {
 910         return;
 911     }
 912     QLIST_REMOVE(vbasedev, next);
 913     vbasedev->group = NULL;
 914     trace_vfio_put_base_device(vbasedev->fd);
 915     close(vbasedev->fd);
 916 }
 917
 918 static int vfio_container_do_ioctl(AddressSpace *as, int32_t groupid,
 919                                    int req, void *param)
 920 {
 921     VFIOGroup *group;
 922     VFIOContainer *container;
 923     int ret = -1;
 924
 925     group = vfio_get_group(groupid, as);
 926     if (!group) {
 927         error_report("vfio: group %d not registered", groupid);
 928         return ret;
 929     }
 930
 931     container = group->container;
 932     if (group->container) {
 933         ret = ioctl(container->fd, req, param);
 934         if (ret < 0) {
 935             error_report("vfio: failed to ioctl %d to container: ret=%d, %s",
 936                          _IOC_NR(req) - VFIO_BASE, ret, strerror(errno));
 937         }
 938     }
 939
 940     vfio_put_group(group);
 941
 942     return ret;
 943 }
 944
 945 int vfio_container_ioctl(AddressSpace *as, int32_t groupid,
 946                          int req, void *param)
 947 {
 948     /* We allow only certain ioctls to the container */
 949     switch (req) {
 950     case VFIO_CHECK_EXTENSION:
 951     case VFIO_IOMMU_SPAPR_TCE_GET_INFO:
 952         break;
 953     default:
 954         /* Return an error on unknown requests */
 955         error_report("vfio: unsupported ioctl %X", req);
 956         return -1;
 957     }
 958
 959     return vfio_container_do_ioctl(as, groupid, req, param);
 960 }